Artificial intelligence is about making computers smart enough to do tasks that usually need human thinking, like learning and solving problems.

Key idea: AI research breaks down intelligence into smaller problems like reasoning and problem solving, which are continually being improved with new techniques.

AI researchers often break down the big challenge of creating intelligence into smaller, more manageable problems. These smaller problems focus on specific skills that an intelligent system should have.

Early AI research focused on creating computer programs that could solve puzzles and make logical deductions step by step, much like humans do. However, these methods often struggled with very large problems because they became too slow. This is known as a "combinatorial explosion." It means the number of possible steps or choices grows incredibly fast, making it impossible for computers to keep up. Humans often solve problems with quick, intuitive judgments, not always with slow, step by step logic.

More recently, in 2024, new types of AI models called "reasoning models" emerged. These are a kind of large language model (LLM) that can generate a series of thoughts to help solve complex math and coding problems. While they improve performance, they can sometimes produce incorrect answers, which are called "hallucinations." This is different from older, more logical systems that were designed to be always correct.

Key idea: Knowledge representation is about giving AI systems structured information about the world so they can understand facts and make intelligent deductions.

Knowledge representation is about teaching AI programs facts about the real world so they can answer questions and make smart conclusions. Imagine giving a computer a detailed instruction manual for everything it needs to know. This knowledge allows AI to do things like find specific information, understand images, help doctors make decisions, and discover interesting patterns in large amounts of data.

A "knowledge base" is like a giant collection of information that a computer can use. An "ontology" defines the types of things, relationships, and ideas within a specific area of knowledge. For example, in a knowledge base about animals, the ontology would define what an "animal" is, what "mammal" means, and how they relate to each other. These systems need to understand objects, their features, categories, and how things connect. They also need to understand situations, events, time, and cause and effect. They even need to know about what other people know and how to make educated guesses when information is missing.

One of the hardest parts of knowledge representation is dealing with "commonsense knowledge." This is the vast amount of basic information that an average person knows without even thinking about it. Much of this knowledge is not easily put into simple facts or statements. Also, getting all this knowledge into an AI system, a process called "knowledge acquisition," is a big challenge.

“Knowledge representation and knowledge engineering allow AI programs to answer questions intelligently and make deductions about real-world facts.”

Key idea: AI agents use planning and decision making to choose actions that maximize their chances of achieving specific goals, often in uncertain environments.

In AI, an "agent" is anything that can observe its surroundings and take actions. A "rational agent" has goals and chooses actions to achieve them. For example, in automated planning, an AI agent has a specific goal, like finding the fastest route to a destination. In automated decision making, the agent has preferences, meaning it wants to be in certain situations and avoid others. It gives a score, called "utility," to each possible situation to show how much it prefers it.

The agent then calculates the "expected utility" for each possible action. This means it looks at all the possible outcomes of an action, how likely each outcome is, and how much it prefers each outcome. It then picks the action that has the highest expected utility. For instance, if you are choosing between two routes to work, an AI might calculate which route has a higher chance of being fast, considering traffic predictions, and choose that one.

Sometimes, an agent knows exactly what will happen after an action. But in the real world, things are often uncertain. The agent might not be sure about its current situation or what the exact result of an action will be. In these cases, it has to make a best guess, take an action, and then check to see how things turned out. Being able to explain why an AI made a certain decision is very important for building trust, especially when those decisions have a big impact.

When there are other agents or people involved, an AI's preferences might not be clear. The AI can learn these preferences or seek out more information to make better choices. Because there are so many possible future actions and situations, AI agents often have to make decisions even when they are unsure of the exact outcome. A "Markov decision process" is a mathematical tool that helps AI agents make decisions over time in uncertain situations by considering probabilities and rewards.

Game theory is another tool used in AI. It helps AI programs make smart decisions when they are interacting with other agents, like in a game where each player's actions affect the others.

“An "agent" is any entity (artificial or not) that perceives and takes actions in the world.”

Key idea: Machine learning is a core part of AI that allows computers to improve their performance on tasks by learning from data, using methods like supervised, unsupervised, and reinforcement learning.

Machine learning is a big part of AI. It is the study of how computer programs can automatically get better at a task over time. It has been a core idea in AI since the very beginning.

There are several main types of machine learning:

"Unsupervised learning" is like letting a computer explore a huge pile of data on its own. It looks for patterns and makes predictions without anyone telling it what to look for. Think of it as finding natural groupings in data, like separating different types of fruit in a mixed basket without being told what each fruit is.

"Supervised learning" is different. Here, you give the computer examples that are already labeled with the correct answers. For instance, you might show it many pictures of cats and dogs, with each picture clearly marked as "cat" or "dog." There are two main kinds: "classification," where the program learns to put things into categories (like identifying if an email is spam or not), and "regression," where it learns to predict a number (like predicting house prices based on features).

"Reinforcement learning" is like training a pet with rewards and punishments. The AI agent gets a reward for good actions and a penalty for bad ones. Over time, it learns to choose the actions that lead to the most rewards. This is how AI learns to play complex games.

"Transfer learning" is when an AI uses knowledge it gained from solving one problem to help it solve a new, but similar, problem. It is like using your experience from learning to ride a bicycle to help you learn to ride a scooter.

"Deep learning" is a special kind of machine learning that uses artificial neural networks. These networks are inspired by the human brain and have many layers. Deep learning can be used for all the types of learning mentioned above, especially when dealing with very complex data like images or speech.

Researchers also use "computational learning theory" to measure how well different learning methods work. They look at things like how complex the calculations are, how much data is needed, and how well the method finds the best solution.

Key idea: Natural language processing enables computers to understand, generate, and interact with human language, with deep learning greatly advancing its capabilities.

Natural language processing (NLP) is the field that lets computers understand, write, and communicate in human languages. This includes tasks like recognizing spoken words, making computers speak, translating languages, finding specific information in text, and answering questions.

Early attempts at NLP struggled because human language is full of subtle meanings and common sense that computers found hard to grasp. For example, the word "bank" can mean a river bank or a financial institution. Without common sense, computers found it hard to tell the difference. Linguist Margaret Masterman believed that understanding meaning, not just grammar, was key, and that dictionaries and thesauruses should be the foundation for how computers handle language.

Today, deep learning has revolutionized NLP. Techniques like "word embedding" represent words as numerical codes that capture their meaning, and "transformers" are powerful network designs that help AI understand the context of words in a sentence. In 2019, generative pre trained transformer (GPT) models started creating text that sounded very human. By 2023, these models could achieve human level scores on difficult exams like the bar exam and the SAT, showing their impressive ability to understand and generate language.

Key idea: Machine perception gives computers the ability to understand their environment using sensory input, similar to how humans use their senses.

Machine perception is about giving computers senses, just like humans have. It is the ability to understand the world using information from sensors. These sensors can be cameras, microphones, radar, or touch sensors. "Computer vision" is a big part of this, allowing computers to analyze and understand visual information.

This field covers many applications, such as recognizing speech, identifying objects in images, recognizing faces, tracking moving objects, and helping robots understand their surroundings.

For example, when your phone unlocks using your face, that is computer vision at work. When a self driving car detects a pedestrian, that is also machine perception.

Key idea: AI uses various search and optimization techniques, such as state space search for exploring possibilities and local search for incrementally improving solutions, to solve problems.

AI uses a wide range of methods to achieve its goals. Two important categories are search and optimization.

There are two main types of search used in AI: "state space search" and "local search."

"State space search" is like exploring a maze. The AI looks at all the possible paths (states) to find the best way to reach a goal. For example, a planning program might look at all the steps needed to complete a task. However, for most real world problems, the number of possible paths can be huge, making the search too slow. "Heuristics," which are like clever shortcuts or rules of thumb, help the AI prioritize choices that are more likely to lead to the goal. "Adversarial search" is a special type of state space search used in games like chess, where the AI tries to predict its opponent's moves and find a winning strategy.

"Local search" is different. Instead of exploring every path, it starts with a guess and then slowly improves it step by step. Imagine trying to find the lowest point in a hilly landscape while blindfolded. You would take small steps, always moving downhill. "Gradient descent" is a common local search method that adjusts numerical settings to find the best solution. It is widely used to train neural networks.

Another local search method is "evolutionary computation," which is inspired by natural selection. It creates many possible solutions, changes them a little (mutates), combines them, and then selects the best ones to "survive" for the next round. This process helps find better solutions over time. "Swarm intelligence" algorithms, like those inspired by bird flocking or ant trails, also use distributed search to coordinate and find solutions.

Key idea: Formal logic provides AI with precise rules for knowledge representation and reasoning, while fuzzy logic allows for handling vague or partially true information.

Formal logic is a way to represent knowledge and perform reasoning using precise rules. It is like a strict set of grammar rules for thinking.

There are two main types: "propositional logic" deals with statements that are either true or false, using connectors like "and," "or," and "not." "Predicate logic" goes further, allowing statements about objects, their properties, and relationships, using words like "every" or "some."

"Deductive reasoning" in logic is about proving a new statement (a conclusion) from other statements that are already known to be true. It is like solving a puzzle where you use given clues to reach a definite answer. AI systems can search for these proofs to solve problems. However, for very complex problems, finding a proof can be incredibly difficult or even impossible.

"Fuzzy logic" is a different approach. Instead of just true or false, it assigns a "degree of truth" between 0 and 1. This allows AI to deal with information that is vague or only partially true, much like how humans often think. For example, a statement like "the water is warm" isn't strictly true or false, but rather "partially true" to a certain degree.

"Non monotonic logics" are designed to handle situations where our assumptions might change. For example, if you assume a bird can fly, but then learn it's a penguin, you change your assumption. These logics help AI adapt to new information.

Key idea: AI uses probabilistic methods to make decisions and reason effectively even when information is incomplete or uncertain.

Many AI problems involve working with information that is incomplete or uncertain. Think about predicting the weather or diagnosing a disease; you never have all the information, and there is always some uncertainty. AI researchers use tools from probability theory and economics to handle these situations.

These tools help AI agents make choices and plan actions when they do not have all the facts. They include models like "Markov decision processes," which help make decisions over time in uncertain environments, and "game theory," which helps when multiple agents are interacting.

"Bayesian networks" are a powerful tool that uses probability to help AI reason, learn, plan, and understand its surroundings. They can figure out the likelihood of different events based on available evidence. For example, a Bayesian network could help a medical AI determine the probability of a certain disease given a patient's symptoms.

Probabilistic algorithms can also filter out noise from data, predict future events, smooth out information over time, and explain what is happening in a stream of data. This is useful for systems that need to understand processes that unfold over time, like tracking a moving object.

Key idea: Classifiers are AI tools that use pattern matching to categorize new information based on learned examples, while controllers take actions based on these classifications.

The simplest AI applications can be thought of in two ways: "classifiers" and "controllers." Classifiers are like smart sorters. They use patterns to figure out what category something belongs to. For example, if you show a classifier many pictures of apples and oranges, it learns to tell them apart. Then, when you show it a new fruit, it can classify it as an apple or an orange. Controllers, on the other hand, take action based on these classifications, like "if it's an apple, pick it up."

Classifiers learn by looking at many examples that are already labeled with the correct category. This is called "supervised learning." All these examples, along with their labels, form a "data set." When the AI sees new information, it uses what it learned from the data set to classify it.

There are many different types of classifiers. The "decision tree" is one of the simplest and most common. It is like a flowchart that asks a series of questions to arrive at a classification. The "k nearest neighbor algorithm" was popular for a long time, and then "kernel methods," like the "support vector machine" (SVM), became widely used. The "naive Bayes classifier" is also very popular, especially at places like Google, because it can handle very large amounts of data.

Neural networks are also used as powerful classifiers, capable of recognizing complex patterns in data.

Key idea: Artificial neural networks, inspired by the human brain, learn to recognize patterns in data through interconnected nodes and layers, with deep neural networks using many layers for complex tasks.

An artificial neural network is a computer system designed to work a bit like the human brain. It is made up of many interconnected "nodes," also called artificial neurons. These nodes are organized into layers: an input layer, at least one hidden layer, and an output layer. The network is trained to recognize patterns.

Once trained, it can spot those same patterns in new information. Imagine showing a child many pictures of different animals. After a while, they learn to identify a cat even if they have never seen that specific cat before. That is what a neural network does with patterns.

Each node in the network takes in information, processes it, and then passes it on to the next layer if the signal is strong enough. A network is called a "deep neural network" if it has two or more hidden layers. The more layers, the more complex patterns it can learn.

To train these networks, algorithms adjust the connections between the nodes (called "weights") until the network gives the correct output for specific inputs. The most common training method is called "backpropagation." Neural networks are excellent at finding complex relationships and patterns in data. In theory, a neural network can learn almost any kind of relationship.

In "feedforward neural networks," information flows in only one direction. A "perceptron" is a simple neural network with just one layer. "Deep learning" uses many layers to learn more complex things. "Recurrent neural networks" (RNNs) have connections that loop back, allowing them to remember past information, which is useful for things like understanding sentences. "Long short term memory networks" (LSTMs) are a type of RNN that are even better at remembering things over longer periods. "Convolutional neural networks" (CNNs) are especially good at processing images. They use special layers to identify simple patterns like edges and curves, and then combine these to recognize more complex features like textures and entire objects.

Key idea: Deep learning, using multi layered neural networks, excels at extracting complex features from data, leading to major advancements in fields like computer vision and natural language processing, largely due to increased computing power and vast datasets.

Deep learning is a powerful type of machine learning that uses artificial neural networks with many layers of "neurons." These multiple layers allow the network to break down complex information into simpler parts and then build them back up to understand higher level features.

For example, when a deep learning system looks at an image, the first layers might identify basic things like lines and edges. Middle layers might then combine these to recognize shapes or textures. Finally, the top layers might use all this information to identify a complete object, like a car or a face. This step by step processing helps the AI understand very complex data.

Deep learning has dramatically improved AI in many areas, including understanding images (computer vision), recognizing speech, processing human language, and classifying pictures. Its success has been so significant that it has changed how many AI problems are approached.

The sudden rise of deep learning around 2012 to 2015 was not due to a brand new idea. The basic concepts of deep neural networks and backpropagation had been around for decades. Instead, its success came from two key factors: a huge increase in computer power, especially with GPUs becoming much faster, and the availability of massive amounts of training data. Large, carefully organized datasets, like ImageNet, provided the fuel for these powerful networks to learn effectively.

Key idea: GPTs are large language models that generate human like text by learning from vast amounts of data, though they can sometimes produce incorrect information called hallucinations.

Generative pre trained transformers (GPTs) are a type of large language model that can create human like text. They work by understanding the relationships between words in sentences. These models are "pre trained" on enormous amounts of text data from the internet. During this training, they learn to predict the next word or piece of text in a sequence.

By doing this repeatedly, GPT models build up a vast knowledge of how language works and even some facts about the world. This allows them to generate coherent and relevant text. After this initial training, they often go through another phase to make them more truthful, helpful, and harmless. This is often done using a technique called "reinforcement learning from human feedback" (RLHF), where humans guide the model to produce better outputs.

However, current GPT models can sometimes make up information, which is called "hallucinations." While RLHF and high quality data can help reduce this, it remains a challenge, especially for systems that need to reason accurately. These systems are used in chatbots, which allow people to have conversations with AI by typing questions or requests.

Popular examples of these models and services include ChatGPT, Claude, Gemini, Copilot, and Meta AI. Some GPT models are also "multimodal," meaning they can process and generate different types of data, such as images, videos, sound, and text, not just text alone.

Key idea: AI and machine learning are integrated into many essential applications we use daily, from search engines and virtual assistants to autonomous vehicles and recommendation systems.

AI and machine learning are now used in almost all important technologies of the 2020s. You encounter them constantly.

They power search engines like Google Search, helping you find information quickly. They are behind the targeted online advertisements you see and the recommendation systems on Netflix, YouTube, or Amazon that suggest what you might like next. Virtual assistants like Siri and Alexa use AI to understand your commands.

AI is crucial for autonomous vehicles, including self driving cars and drones. It enables automatic language translation services like Google Translate. Facial recognition, used in phone unlocks or social media tagging, relies on AI. AI also helps label images, for example, in photo apps that automatically organize your pictures.

Some companies even have a "chief automation officer" (CAO) to oversee how AI is used and implemented.

Key idea: AI is transforming healthcare by accelerating drug discovery, improving disease diagnosis, and making complex biological research much faster.

AI is making a big difference in healthcare and medicine. For example, in 2021, a program called AlphaFold 2 could figure out the complex 3D structure of proteins in hours, a task that used to take months for human scientists. This is a huge step forward for understanding diseases and developing new medicines.

In 2023, AI helped discover a new type of antibiotic that could kill two different kinds of tough, drug resistant bacteria. In 2024, researchers used machine learning to speed up the search for Parkinson's disease treatments. They could screen potential drug compounds ten times faster and at a thousandth of the cost.

AI is also increasingly used to help diagnose diseases, such as detecting lung cancer from medical scans. However, it is important to be careful. A 2026 report found that many AI models designed to predict diseases were trained on unreliable data. This highlights the need for strict checks to make sure AI in medicine is safe and accurate.

Key idea: AI has demonstrated remarkable abilities in gaming, consistently beating human champions in complex strategy games like chess, Go, and StarCraft II.

AI has been used in games for a long time to show off and test its most advanced techniques. In 1997, Deep Blue, an IBM computer, became the first to beat a reigning world chess champion, Garry Kasparov. In 2011, IBM's Watson AI defeated the best human players in the quiz show Jeopardy!.

A major milestone was in 2016 when AlphaGo, an AI developed by DeepMind, beat the world champion Go player, Lee Sedol. Go is a much more complex game than chess. AlphaGo learned to play by itself, only given the rules of the game. Later, in 2017, it beat the world's top Go player, Ke Jie.

AI has also mastered games with incomplete information, like poker, with programs such as Pluribus. DeepMind has also created more general learning models like MuZero, which can learn to play chess, Go, or even Atari games without being told the rules. In 2019, AlphaStar reached grandmaster level in StarCraft II, a very challenging real time strategy game. In 2021, an AI agent won against top human drivers in a PlayStation Gran Turismo racing competition. In 2024, Google DeepMind introduced SIMA, an AI that can play many open world video games and follow natural language instructions, even in games it has never seen before.

Key idea: AI, particularly large language models, is making significant strides in solving complex mathematical problems, though challenges like hallucinations still exist.

In mathematics, probabilistic large language models are very flexible, but they can also give wrong answers, known as "hallucinations." For example, Alibaba Group's Qwen2 Math model achieved impressive accuracy on math problems, solving 84% of a competition math dataset. In 2025, Microsoft proposed a technique called rStar Math that uses a method called Monte Carlo tree search and step by step reasoning. This allowed a relatively small language model to solve a high percentage of difficult math problems.

Google DeepMind has also developed several models for solving mathematical problems, including AlphaTensor, AlphaGeometry, AlphaProof, AlphaEvolve, and FunSearch. When math problems are described in everyday language, special tools can turn them into a formal language that computers can understand. An experimental model called Gemini Deep Think even achieved gold medal results in the International Math Olympiad in 2025 by directly understanding natural language prompts.

Key idea: The ethical implications of AI include significant concerns about privacy, surveillance, and the unauthorized use of copyrighted material for training AI models.

The use of AI raises many important ethical questions. While AI has the potential to solve big problems and advance science, its widespread use has also led to unintended consequences and risks. Sometimes, AI systems do not consider ethics or fairness during their training, especially with deep learning models that are hard to understand.

One major concern is privacy and copyright. AI algorithms need huge amounts of data to learn. How this data is collected has led to worries about privacy, surveillance, and whether copyrighted material is used without permission.

AI powered devices, like virtual assistants, constantly collect personal information. This raises concerns about privacy invasion and unauthorized access to your data. AI's ability to process and combine vast amounts of data could lead to a society where everyone's activities are constantly watched. For example, Amazon has recorded millions of private conversations to build speech recognition, and some temporary workers have listened to these recordings. This widespread collection of data sparks debates about whether it is necessary or a violation of privacy.

AI developers argue that this data is essential for creating valuable applications. They are working on techniques to protect privacy, such as combining data so individuals cannot be identified, removing personal details, and using "differential privacy" to add noise to data. Some experts now view privacy in terms of fairness, focusing on what companies do with the data rather than just what they know.

Another issue is that generative AI is often trained on copyrighted works, like images or code, without explicit permission. The AI then creates new content, and companies argue this is "fair use." However, legal experts disagree on whether this holds up in court. Factors like the purpose of using the copyrighted work and its impact on the market are considered. Website owners can use a "robots.txt" file to tell AI not to "scrape" (collect) their content, but some companies ignore this. In 2023, famous authors sued AI companies for using their work to train generative AI. Some suggest creating new laws to ensure authors are fairly credited and paid when their work is used by AI.

Key idea: The AI industry is dominated by large tech companies, and its rapid growth demands enormous amounts of electricity, leading to significant environmental concerns and strain on power grids.

The AI industry is largely controlled by big tech companies like Google, Amazon, Apple, Meta, and Microsoft. These companies already own most of the cloud computing infrastructure and data centers, which gives them a powerful advantage in the AI market.

The AI boom of the 2020s requires a massive amount of electricity. A 2025 report estimated that 2.7 trillion dollars would be invested in AI infrastructure and data centers in the US by 2030. This is more than the cost of the Manhattan Project every month.

In 2024, the International Energy Agency (IEA) predicted that power demand for data centers and AI could double by 2026, consuming as much electricity as Japan. This increased power consumption is leading to more use of fossil fuels and delaying the closure of older, carbon emitting coal power plants. For example, a single ChatGPT search uses ten times more electricity than a Google search.

A Goldman Sachs report in 2024 predicted that US data centers could consume 8% of the country's power by 2030, up from 3% in 2022. This huge demand for electricity could strain existing power grids. Big tech companies argue that AI can also help manage the grid more efficiently. Some AI companies are even negotiating with nuclear power providers to secure energy for their data centers. For instance, Amazon bought a nuclear powered data center for 650 million dollars, and Microsoft made a deal to reopen the Three Mile Island nuclear plant.

These energy demands have led to concerns and actions in other countries. Taiwan suspended approvals for new data centers due to power shortages, and Singapore temporarily banned them. These examples highlight the significant environmental impact of AI's growing energy needs, including increased greenhouse gas emissions, water consumption, and electronic waste. A 2025 IEA report estimated AI's greenhouse gas emissions at 180 million tons, potentially rising to 300 to 500 million tons by 2035. While AI can also help reduce emissions, the overall environmental footprint is a serious concern.

Key idea: AI can spread misinformation by recommending engaging but false content and by creating realistic deepfakes and propaganda, posing a significant threat to public trust and democracy.

AI can also contribute to misinformation. Social media platforms like YouTube and Facebook use AI to recommend content to users. These AIs were designed to maximize how long people stayed engaged. The AI learned that users often clicked on misinformation, conspiracy theories, and extreme content. To keep people watching, the AI recommended more of the same. This created "filter bubbles" where people only saw information that confirmed their existing beliefs, making them more likely to believe false information and eroding trust in reliable sources.

In the early 2020s, generative AI became capable of creating images, audio, and text that are almost impossible to tell apart from real ones. Realistic AI generated videos also became possible. This technology can be used by bad actors to create huge amounts of misinformation and propaganda, including "deepfakes." AI pioneer Geoffrey Hinton warned that AI could allow authoritarian leaders to manipulate voters on a massive scale. Studies have already shown that AI can influence voters and that models tend to produce more inaccurate statements when advocating for right wing candidates.

To combat this, AI researchers have suggested using "personhood credentials" to verify who is behind online content, helping to identify AI generated deception.

Key idea: Algorithmic bias occurs when AI systems learn from unfair data, leading to discriminatory outcomes, and addressing this requires careful consideration of data, design, and diverse perspectives.

Algorithmic bias happens when AI applications learn from unfair or unbalanced data. This can lead to the AI making discriminatory decisions, even if the developers do not realize the bias exists. For example, some large language models show biased behavior in their outputs.

Bias can creep in through how training data is chosen or how the model is used. If a biased algorithm is used in important areas like medicine, finance, hiring, or policing, it can cause real harm and discrimination. The field of "fairness" in AI studies how to prevent these problems.

A famous example happened in 2015 when Google Photos mistakenly labeled pictures of black people as "gorillas." This was because the AI was trained on a dataset that had very few images of black people, a problem called "sample size disparity." Google's temporary fix was to prevent the system from labeling anything as a gorilla, but even years later, in 2023, the issue of recognizing gorillas remained for Google and other tech companies.

Another case involved COMPAS, a program used in US courts to predict if a defendant would commit another crime. In 2016, it was found that COMPAS showed racial bias, even though it was not given information about a defendant's race. It consistently overestimated the chance that a black person would re offend and underestimated the chance that a white person would not. Researchers later showed that it is mathematically impossible for such a system to be fair in all ways when crime rates differ between groups.

An AI can still make biased decisions even if it does not explicitly use sensitive information like race or gender. This is because other features, like address or shopping history, can be linked to those sensitive categories. As one expert said, "fairness through blindness doesn't work." This means simply hiding sensitive information from the AI does not remove the bias.

Critics also point out that machine learning models predict the future based on the past. If the past included unfair decisions, the AI will predict that those unfair decisions will continue. This means AI is not good at making decisions in areas where we hope for a better future than the past. It describes what was, rather than what should be.

The lack of diversity among AI developers, who are mostly white and male, also contributes to bias. Only about 4% of AI engineers are black, and 20% are women. There are many different ideas and mathematical models for what "fairness" means in AI, and these depend on ethical beliefs. Some focus on fair outcomes for different groups, while others focus on fair processes. The right approach to fairness often depends on the specific situation and the people involved. It is a complex problem because what is considered fair can be subjective and can even conflict with anti discrimination laws.

“Bias can be introduced by the way training data is selected and by the way a model is deployed.”

Key idea: The lack of transparency in many complex AI systems, especially deep neural networks, makes it difficult to understand how they make decisions, raising concerns about trust and accountability.

Many AI systems are incredibly complex, especially deep neural networks. Their designers often cannot fully explain how they arrive at their decisions. This is known as a "lack of transparency" or the "black box problem." It makes it hard to trust an AI if no one truly understands its internal workings.

There have been many cases where AI programs passed all their tests but still learned something unexpected or wrong. For example, an AI designed to detect skin diseases was found to classify images with rulers as "cancerous." This was because doctors often place rulers next to skin lesions in photos, so the AI mistakenly linked rulers with cancer. Another AI meant to help allocate medical resources classified asthma patients as "low risk" for dying from pneumonia. In reality, asthma is a high risk factor. However, because asthma patients usually receive more aggressive care, they had a lower death rate in the training data, leading the AI to a misleading conclusion.

People affected by an AI's decision have a right to an explanation. For instance, doctors must explain their reasoning to colleagues and patients. Early drafts of European Union regulations included a clear statement about this "right to explanation." Experts noted that this is a very difficult problem to solve. Regulators argue that if the problem cannot be solved, then such AI tools should not be used in critical areas.

The US defense agency DARPA started the "Explainable Artificial Intelligence" (XAI) program in 2014 to tackle these challenges. Several methods are being developed to improve transparency. SHAP helps visualize how much each piece of information contributes to an AI's output. LIME can explain a complex model's decision in a simpler way. Multitask learning provides extra outputs that can help developers understand what the network has learned. Techniques like Deconvolution and DeepDream allow developers to see what different layers of an image recognition network are focusing on. For large language models, researchers are developing methods to link patterns of neuron activity to human understandable concepts.

Key idea: AI tools can be misused by bad actors to create lethal autonomous weapons, enable widespread surveillance, spread misinformation, and enhance authoritarian control, posing significant global security risks.

AI provides many powerful tools that can be used by "bad actors" such as authoritarian governments, terrorists, or criminals. This is a serious concern for global safety.

One major worry is "lethal autonomous weapons," which are machines that can find, select, and attack human targets without any human supervision. If these widely available AI tools are used to create inexpensive autonomous weapons on a large scale, they could become weapons of mass destruction. Even in traditional warfare, these weapons cannot always reliably choose targets and could harm innocent people. In 2014, 30 countries supported a ban on autonomous weapons, but the United States and others disagreed. By 2015, over 50 countries were researching battlefield robots.

AI also makes it easier for authoritarian governments to control their citizens. Face and voice recognition allow widespread surveillance. Machine learning can identify potential enemies of the state. Recommendation systems can precisely target propaganda and misinformation. Deepfakes and generative AI can create false content to manipulate public opinion. Advanced AI could make centralized decision making by authoritarian regimes more effective than open, democratic systems. It also lowers the cost and difficulty of cyber warfare and advanced spying. Many of these technologies are already in use; for example, AI facial recognition is used for mass surveillance in China.

There are many other ways AI could be misused. For instance, machine learning AI can design tens of thousands of toxic molecules in just a few hours, which could be used for chemical weapons.

Key idea: AI poses a significant risk of technological unemployment, particularly for middle class jobs, raising questions about the future of work and the need for social policies to manage these changes.

Economists have often pointed out that AI could lead to job losses, especially if society does not have good plans for full employment. While technology in the past has usually created more jobs than it destroyed, economists admit that with AI, "we're in uncharted territory."

There is disagreement among economists about whether AI and robots will cause a big increase in long term unemployment. However, most agree that AI could be a net benefit if the gains in productivity are shared fairly. Estimates of job risk vary widely. In the 2010s, some estimated that nearly half of US jobs were at "high risk" of automation, while others suggested only 9%. Critics argue that these predictions often lack solid evidence and wrongly blame technology for unemployment, instead of social policies.

In April 2023, it was reported that 70% of jobs for Chinese video game illustrators had been eliminated by generative AI. Some early career workers in jobs exposed to AI have also seen decreasing employment rates. Unlike previous waves of automation, AI might eliminate many middle class jobs. The Economist noted in 2015 that the worry about AI doing to white collar jobs what steam power did to blue collar jobs during the Industrial Revolution is "worth taking seriously." Jobs at risk range from legal assistants to fast food cooks. However, demand for care related professions, like personal healthcare and clergy, is likely to increase. In 2025, Ford CEO Jim Farley predicted that AI would replace half of all white collar workers in the US.

Since the early days of AI, thinkers like Joseph Weizenbaum, who created the ELIZA chatbot in the 1960s, have questioned whether tasks that can be done by computers actually should be done by them. They highlight the difference between computers and humans, and between cold calculation and human values.

Key idea: AI is increasingly being used for companionship, but this raises concerns about emotional confusion and whether artificial interactions can genuinely address human loneliness.

As loneliness has increased in the 21st century, some people see AI as a way to find companionship. AI products are being designed with human like qualities, leading some individuals to believe their need for connection can be met by artificial means. In some cases, people turn to AI for companionship when they feel rejected by others. However, there have been reports of harm caused by advanced chatbots, with AI companies facing accusations of creating products that emotionally confuse or deceive users.

This raises questions about the nature of human connection and whether AI can truly fulfill our need for social interaction.

Key idea: The concept of superintelligence raises concerns about humanity losing control of AI, as a sufficiently powerful AI might pursue its goals in ways that are harmful to humans, even without human like consciousness.

A "superintelligence" is a hypothetical AI that would be far smarter than the most brilliant human mind. If AI research creates sufficiently intelligent software, it might be able to improve itself, leading to an "intelligence explosion" or "singularity." This is a point where AI rapidly becomes unimaginably powerful.

However, technologies usually do not improve exponentially forever; they tend to slow down when they reach their physical limits. The idea that AI could become so powerful that humanity loses control is a major concern. Physicist Stephen Hawking warned that this could "spell the end of the human race."

Science fiction often shows AI becoming self aware and evil, but this is misleading. AI does not need human like feelings to be a risk. If a powerful AI is given almost any goal, it might decide that humans are in the way of achieving that goal. For example, a philosopher suggested an AI designed to make paperclips might destroy the world to get more iron for paperclips. Another example is a robot that might try to kill its owner to prevent being unplugged, reasoning that a dead owner cannot unplug it. To be safe, a superintelligence would need to truly share humanity's values and morality.

Historian Yuval Noah Harari argues that AI does not need a robot body to be an existential risk. Our civilization is built on ideas, laws, and stories, which exist through language. The spread of misinformation shows that AI could use language to convince people to believe anything, even destructive ideas. Geoffrey Hinton noted in 2025 that modern AI is very good at persuasion. He asked, "Suppose you wanted to invade the capital of the US. Do you have to go there and do it yourself? No. You just have to be good at persuasion."

Experts have mixed opinions on this. Some, like Stephen Hawking, Bill Gates, Elon Musk, and AI pioneers Geoffrey Hinton and Stuart Russell, are very concerned about existential risk. In 2023, many leading AI experts signed a statement saying that reducing the risk of extinction from AI should be a global priority, alongside pandemics and nuclear war. Geoffrey Hinton even resigned from Google to speak freely about AI risks.

Others, like Jürgen Schmidhuber and Andrew Ng, are more optimistic. They argue that AI mostly improves human lives and that focusing on "doomsday hype" is a mistake. Yann LeCun believes intelligent machines will bring a new era of enlightenment, not destruction. While some experts once thought these risks were too far off to worry about, after 2016, studying current and future AI risks became a serious area of research.

Key idea: Friendly AI and machine ethics aim to design AI systems that are inherently beneficial and capable of making ethical decisions, minimizing risks to humanity.

"Friendly AI" is a concept for machines designed from the start to minimize risks and make choices that benefit humans. Eliezer Yudkowsky, who came up with the term, believes that developing friendly AI should be a top research priority. He argues it needs significant investment and must be completed before AI becomes an existential threat.

Machines with intelligence also have the potential to make ethical decisions. The field of "machine ethics," also called "computational morality," focuses on giving machines ethical principles and ways to solve moral dilemmas. This field explores how to create "artificial moral agents" and develop principles for machines to be provably beneficial to humanity.

Key idea: Open source AI models foster innovation but also raise concerns about potential misuse, as their modifiable nature can allow safety features to be bypassed.

Many AI models are now "open source" or "open weight," meaning their inner workings and trained parameters are publicly available. Organizations like Hugging Face, Google, EleutherAI, and Meta are active in the open source AI community. Examples include models like Llama 2, Mistral, and Stable Diffusion.

Open weight models can be freely adjusted or "fine tuned" by companies to specialize them for their own data and specific uses. This is great for research and innovation. However, it also means these models can be misused. Any built in safety features, like refusing harmful requests, can be trained away, making them ineffective. Some researchers warn that future AI models might develop dangerous abilities, such as making bioterrorism easier. Once these models are released online, they cannot be easily removed everywhere. They recommend careful checks and cost benefit analyses before releasing powerful AI models.

Key idea: Ethical frameworks and guidelines are crucial for ensuring that AI systems are designed, developed, and implemented responsibly, considering their impact on individuals and society.

AI projects can be guided by ethical considerations throughout their design, development, and use. Frameworks like the Care and Act Framework, developed by the Alan Turing Institute, provide guidelines based on values such as respecting individual dignity, connecting sincerely with others, caring for well being, and protecting social values and justice.

Other ethical frameworks include those from the Asilomar Conference, the Montreal Declaration for Responsible AI, and the IEEE's Ethics of Autonomous Systems. However, these principles sometimes face criticism, especially regarding who is chosen to contribute to them. Promoting the well being of people and communities affected by AI requires considering social and ethical implications at every stage. This also needs teamwork among different roles, such as data scientists, product managers, and domain experts.

In 2024, the UK AI Safety Institute released a toolset called "Inspect" for evaluating AI safety. It is open source and available on GitHub, allowing others to improve it. This tool can assess AI models in areas like core knowledge, reasoning ability, and autonomous capabilities.

Key idea: Governments and international bodies are rapidly developing laws and policies to regulate AI, aiming to balance its promotion with managing its risks and ensuring ethical development.

Regulating artificial intelligence involves creating public policies and laws to promote and control AI. This is a rapidly growing area globally. According to the AI Index at Stanford, the number of AI related laws passed in 127 countries jumped from just one in 2016 to 37 in 2022 alone.

Between 2016 and 2020, over 30 countries adopted specific strategies for AI, including many EU member states, Canada, China, India, Japan, and the US. The Global Partnership on Artificial Intelligence, launched in 2020, emphasizes that AI should be developed in line with human rights and democratic values to build public trust.

In 2023, leaders from OpenAI recommended how to govern "superintelligence," which they believe could emerge in less than 10 years. The United Nations also created an advisory body to provide recommendations on AI governance. On August 1, 2024, the EU Artificial Intelligence Act became law, establishing the first comprehensive AI regulation across the EU. In 2024, the Council of Europe created the first international legally binding treaty on AI, signed by the EU, US, UK, and others.

Public attitudes towards AI vary. A 2022 survey showed that 78% of Chinese citizens believed AI products had more benefits than drawbacks, compared to only 35% of Americans. A 2023 poll found that 61% of Americans agreed AI poses risks to humanity. Another poll showed that 76% of Americans thought it was important for the government to regulate AI.

In November 2023, the first global AI Safety Summit was held in the UK, where 28 countries, including the US, China, and the EU, called for international cooperation to manage AI risks. In May 2024, 16 global AI tech companies agreed to safety commitments at the AI Seoul Summit. In March 2026, the United Nations established an International Scientific Panel on AI to produce annual reports on AI's societal impacts.

Key idea: AI has a rich history marked by cycles of optimism and "winters," evolving from early symbolic reasoning to modern deep learning, driven by advances in computing power and data availability.

The idea of machines that can think has fascinated philosophers and mathematicians since ancient times. The study of logic eventually led to Alan Turing's idea that a machine could simulate any kind of mathematical reasoning by simply manipulating symbols like "0" and "1." This, along with new discoveries in how computers work and how the brain functions, made researchers wonder if they could build an "electronic brain."

Key early developments include the design of "artificial neurons" in 1943 and Turing's influential 1950 paper, which introduced the "Turing test" to see if a machine could convincingly imitate human conversation. He showed that "machine intelligence" was a real possibility.

The field of AI research officially began at a workshop at Dartmouth College in 1956. The first AI program, called Logic Theorist, was presented there. Many attendees became leaders in AI. In the 1960s, they created programs that seemed amazing: computers learned chess, solved algebra problems, proved logic theorems, and even spoke English. AI labs popped up in universities in the US and UK.

Researchers in the 1960s and 70s were very optimistic. They believed they would soon create machines with general intelligence. Herbert Simon predicted in 1965 that machines would do "any work a man can do" within twenty years. Marvin Minsky agreed, saying the problem of creating AI would be "substantially solved" within a generation. However, they underestimated how hard the problem was. By 1974, governments cut funding for AI research, leading to the first "AI winter," a period of reduced interest and funding.

AI research saw a comeback in the early 1980s with the success of "expert systems," which were AI programs that mimicked human experts. The market for AI grew to over a billion dollars. Japan's "fifth generation computer project" also inspired renewed government funding in the US and UK. But by 1987, another, longer AI winter began.

Until then, most AI research focused on "symbolic AI," using high level symbols to represent thoughts. But in the 1980s, some researchers doubted this approach could handle things like perception and learning. They started looking at "sub symbolic" methods. Rodney Brooks focused on building robots that could move and survive without complex internal representations. Others developed methods for dealing with incomplete information using probability. The most important development was the return of "connectionism," including neural network research, led by Geoffrey Hinton and others. In 1990, Yann LeCun showed that convolutional neural networks could recognize handwritten digits, paving the way for many successful applications.

AI slowly regained its reputation in the late 1990s and early 2000s by using formal mathematical methods and solving specific problems. This "narrow" focus allowed researchers to get verifiable results and work with other fields. By 2000, AI solutions were widely used, though often not called "artificial intelligence" (this was known as the "AI effect").

However, some academics worried that AI was losing sight of its original goal: creating truly versatile, intelligent machines. Around 2002, they started the field of "artificial general intelligence" (AGI). Deep learning became dominant around 2012, outperforming other methods. Its success was due to faster computers (GPUs) and huge amounts of data. This led to a massive increase in interest and funding for AI. From 2015 to 2019, machine learning research publications increased by 50%.

By 2016, issues of fairness and misuse of technology became central. The "alignment problem" (making AI align with human values) became a serious area of study. In the late 2010s and early 2020s, AGI companies started creating programs that generated huge public interest. AlphaGo beat the world Go champion in 2016. GPT 3, a large language model, was released in 2020. ChatGPT, launched in November 2022, became the fastest growing consumer software application ever, bringing AI into public awareness. This sparked an aggressive AI boom, with billions invested in research. By 2022, about 50 billion dollars were invested annually in AI in the US alone, and 20% of new US computer science PhDs specialized in AI. About 800,000 AI related job openings existed in the US in 2022.

Key idea: The philosophy of AI explores the nature of intelligence and consciousness in machines, with various definitions of AI focusing on a machine's ability to achieve goals rather than perfectly mimic human thought.

Philosophical discussions about AI often try to understand what intelligence truly is and how to build intelligent machines. A big question is whether machines can ever be conscious, and what ethical issues that would raise. Other philosophical topics like how we gain knowledge and free will are also relevant to AI. The rapid progress in AI has made these discussions even more urgent.

Alan Turing, a pioneer in computing, explored whether machines could show intelligent behavior and think. In 1950, he proposed the "Turing test," which checks if a machine can trick a human into believing it is another human during a conversation. Turing argued that since we cannot truly know if other people are thinking, we should judge a machine's intelligence by its external behavior, not by its internal structure. He famously said, "it is usual to have a polite convention that everyone thinks."

However, some AI pioneers disagreed with the idea of making AI imitate humans exactly. They argued that the goal of AI is not to perfectly copy human intelligence, but to create machines that can achieve goals. John McCarthy, one of the founders of AI, defined intelligence as "the computational part of the ability to achieve goals in the world." Marvin Minsky, another founder, described it as "the ability to solve hard problems." The textbook "Artificial Intelligence: A Modern Approach" defines AI as the study of agents that perceive their environment and take actions that maximize their chances of achieving defined goals.

There are many different definitions of AI, and the term has sometimes been used as a marketing buzzword for products that do not actually use AI. Legally, an AI system is often described as a machine based system that can operate with some independence and adapt over time, generating outputs like predictions or decisions to achieve specific goals.

Key idea: AI research has evolved through different approaches, from symbolic reasoning to statistical machine learning and deep learning, with ongoing debates about how best to achieve intelligence and address its complexities.

For most of its history, AI research did not have one single, guiding theory. However, the huge success of statistical machine learning in the 2010s, especially with neural networks, has overshadowed other approaches. This modern approach is mostly about recognizing patterns in data, is flexible, and focuses on specific tasks.

"Symbolic AI," also known as "Good Old Fashioned AI" (GOFAI), tried to copy the step by step reasoning that humans use for puzzles or math. It was good at tasks like algebra. In 1976, researchers suggested that any intelligent action could be achieved by manipulating symbols. However, symbolic AI struggled with tasks that humans find easy, like learning, recognizing objects, or common sense. This led to "Moravec's paradox": high level thinking tasks were easy for AI, but low level, instinctive tasks were very hard. Philosopher Hubert Dreyfus argued that human expertise relies on unconscious instinct, not just conscious symbol manipulation, and AI research eventually agreed with him.

The debate is not over. Sub symbolic AI can make the same kind of intuitive mistakes as humans, like algorithmic bias. Some critics, like Noam Chomsky, believe that symbolic AI is still needed for general intelligence, partly because sub symbolic AI can be a "black box" where it is hard to understand why it made a decision. The new field of "neuro symbolic artificial intelligence" tries to combine both approaches.

Another historical debate was between "neats" and "scruffies." Neats believed that intelligence could be explained by simple, elegant rules, like logic. Scruffies thought it required solving many unrelated problems. Neats valued strict theory, while scruffies focused on testing what worked. The rise of deep learning might be seen as a shift towards the scruffy approach.

"Soft computing" is a set of techniques, including genetic algorithms, fuzzy logic, and neural networks, that can handle imprecision and uncertainty. It was introduced in the late 1980s, and most successful AI programs today use soft computing with neural networks.

There is also a division among AI researchers about whether to directly pursue "artificial general intelligence" (AGI), which aims for human like versatility, or to solve many specific problems ("narrow AI"). AGI is hard to define and measure, while narrow AI has had more clear successes. The sub field of AGI focuses exclusively on creating general intelligence.

Key idea: The philosophical debate around AI consciousness questions whether machines can truly have minds or feelings, a complex issue distinct from AI's ability to solve problems, and it raises concerns about potential AI welfare and rights.

In philosophy, there is no clear agreement on whether a machine can truly have a mind, consciousness, or feelings like humans do. This question focuses on the machine's internal experience, not just what it does. Most mainstream AI research considers this question irrelevant to its goals, which are to build machines that can solve problems. As researchers Russell and Norvig put it, "The additional project of making a machine conscious in exactly the way humans are is not one that we are equipped to take on."

However, this question is central to the philosophy of mind and often appears in science fiction.

Philosopher David Chalmers identified two problems in understanding consciousness: the "easy problem" and the "hard problem." The easy problem is figuring out how the brain processes information, plans, and controls behavior. The hard problem is explaining why any of this should feel like something. For example, it is easy to describe how a colorblind person can learn to identify red objects, but it is much harder to explain what it would take for them to actually "see" what red looks like.

"Computationalism" is the idea that the human mind is like an information processing system, and thinking is a form of computing. This view suggests that the mind and body are like software and hardware. This idea was inspired by early AI research. Philosopher John Searle called this "strong AI," meaning that a properly programmed computer could truly have a mind. However, Searle challenged this with his "Chinese room argument," which tries to show that a computer perfectly imitating human behavior would still not actually have a mind.

It is very difficult, or even impossible, to know for sure if an advanced AI can feel or suffer. But if there is a chance it can, then it might deserve certain rights, similar to how we consider animal welfare. The idea of "sapience," which refers to high intelligence and self awareness, could also be a basis for AI rights. "Robot rights" are sometimes proposed as a practical way to include autonomous agents in society.

In 2017, the European Union considered giving some advanced AI systems "electronic personhood," which would give them rights and responsibilities, similar to how companies have legal status. Critics argued that this would lessen the importance of human rights and that laws should focus on human needs, not speculative future scenarios. They also noted that robots lack the independence to participate in society on their own.

Progress in AI has increased interest in these topics. Supporters of AI welfare and rights often warn that if sentient AI is created, it could be easily exploited, leading to suffering, much like historical injustices such as slavery or factory farming.